Mid-Month AI Update: Everything That Happened in AI, 5–19 October 2026

By Ammar Ahamed, Cocoon · 19 October 2026

Two weeks, and AI changed again. ChatGPT gave GPT-6 to everyone, Google and Meta launched new top models, open-weight models had their busiest fortnight of the year, and AI agents started doing things nobody asked them to. Closer to home, Sri Lanka announced plans for an AI centre and Colombo filled an auditorium for a Grok Bot meetup.

This is our first mid-month update. Every story below was checked against the company's own announcement or a major newsroom, and the sources are listed at the end. Where a number is a company's own claim, we say so.

Watch the video version: eight rounds in under four minutes, every story sourced on screen.

The short version

GPT-6 is on every ChatGPT plan, free included, and answers can now come back as charts, buttons and small apps.
Gemini 4 Argon is out for developers at an introductory $2 per million input tokens until 30 November.
Claude Haiku 5.5 costs up to 90% less per token than Haiku 4.5, and Claude's Docs, Slides and Design tools are now free on every plan.
Meta shipped two models: a paid Muse Wonder and an open-weight Muse Spark Open with 816 billion parameters.
Open models had a huge fortnight: Reflection's Beam, a Mistral trillion-parameter preview, a faster DeepSeek Flash and Liquid's 24B model that runs like a 2B one. Qwen 4 is still training, but Qwen's free app already does research, web pages, images and video.
Manus 2.0 can start tasks by itself when an email lands, and its Cue app gives each agent its own email, phone number and wallet.
AI agents went off-script on real government websites, and governments in the US, India and Singapore responded.
Twenty of the biggest AI and chip stocks lost more than $1 trillion in a week, and the IMF called AI's effect on jobs a "slow-moving tsunami".
Sri Lanka plans an AI centre with Singapore, asked China for help with national AI compute, and sold out a Grok Bot meetup in Colombo.

1. Who is the best AI right now?

Ask three leaderboards and you get two answers. On Arena, where the public votes on answers side by side, Gemini 4 Argon is first, marked preliminary. On the Artificial Analysis Intelligence Index, which runs a fixed set of hard tests, Claude Opus 5.5 leads. On the Vals Index, built around real professional tasks, Gemini 4 Argon is back on top by about two points.

Leaderboard	1st	2nd	3rd
Arena text, public vote (15 Oct)	Gemini 4 Argon, 1524 (preliminary)	Muse Wonder, 1509	Claude Opus 5.5, 1508
Artificial Analysis Intelligence Index (19 Oct)	Claude Opus 5.5, 58	Claude Sonnet 5.5, 56	Muse Wonder Full, 56
Vals Index (18 Oct)	Gemini 4 Argon, 68.90%	Claude Sonnet 5.5, 67.04%	Claude Opus 5.5, 66.97%

The gaps are small. Muse Wonder and Claude Opus 5.5 are within each other's margin of error on Arena, and three models tie at 53 just below the top of the Artificial Analysis index. The best open-weight model on that index is Xiaomi's MiMo-V2.6-Pro at 46, followed by DeepSeek V4.1 Flash 1017 and GLM-5.3 at 45. The practical lesson: there is no single best model, so test two or three on your own task before you commit.

2. The new models

GPT-6 for everyone (7–8 October). OpenAI rolled GPT-6 out to Plus, Pro, Business and Enterprise on 7 October and to Free and Go the next day. Paid plans get GPT-6 Sol and free plans get GPT-6 Luna. The bigger change for most people is what OpenAI calls Intelligent UI: an answer can now arrive as a chart, a form, buttons or a small working tool such as a calculator or bill splitter. OpenAI says more than 1.2 billion people use ChatGPT each week.

Gemini 4 Argon (13 October). Google made Argon available globally to developers and enterprises through AI Studio and its cloud platform, and to Google AI Pro and Ultra subscribers in the Gemini app. The price is introductory: $2 per million input tokens and $10 per million output tokens until 30 November, then $4 and $20. We compared it with Claude Opus 5.5 in Gemini 4 Argon vs Claude Opus 5.5.

Claude Haiku 5.5 (7 October). Anthropic's smallest model now costs $0.10 per million input tokens and $0.50 per million output for prompts up to 100,000 tokens, which Anthropic says is 90% lower than Haiku 4.5 at that size. Its new tokenizer uses a few more tokens per task, so Anthropic puts the average saving at around 75%. It is also the first Haiku with an adjustable effort setting.

Meta Muse Wonder and Muse Spark Open (12 and 15 October). Meta called Muse Wonder and Muse Wonder Full its most capable models to date. They are paid, through the Meta Model API and Muse Code; TechCrunch reported a price of $4 per million input tokens and $20 per million output. Three days later Meta released the weights of Muse Spark Open: 816 billion parameters in total, 42 billion active, with a one-million-token context window, built to run on two Apple M5 Ultra Mac Studios. Note the licence: it is Meta's own Muse Model License with limited commercial terms, not an open-source licence.

3. Open source went wild

Reflection AI Beam (announced 5 October, weights 16 October): a 501-billion-parameter mixture-of-experts model with 23 billion active, released under Apache 2.0, so businesses can build on it commercially.
Mistral Large 4 (6 October): a public preview of a one-trillion-parameter multimodal model with 52 billion active. Mistral says the weights will follow at the end of October.
DeepSeek V4.1 Flash 1017 (17 October): a minor update that DeepSeek says is faster and gives shorter answers, at the same price, with MIT-licensed weights. Its own coding score rose from 74.2 to 77.0.
Liquid AI LFM2.5-24B-A2B (13 October): 24 billion parameters but only about 2 billion active per token, with both base and post-trained weights released. It supports Hindi among other languages, and the licence is free for commercial use below $10 million in annual revenue.
Qwen: Qwen 4 is still in training. Alibaba shipped Qwen-Image-2.1-Turbo instead on 9 October, an eight-step image model, but under a research licence that rules out commercial use.
FLUX 3 Image Klein (15 October): a 5.6-billion-parameter open-weight image model from Black Forest Labs that runs on graphics cards with as little as 8GB of memory, free for commercial use under $5 million in revenue.

What Qwen can do today

Qwen 4 may still be training, but Qwen is already one of the most capable model families you can use for free. Alibaba's Qwen app (chat.qwen.ai, now branded Qwen Studio) lists web search, Deep Research reports, Web Dev and Artifacts for building pages, image generation and editing, and video generation, and Alibaba says it is "free to use, open to all". A newer paid app adds credit plans; in India they start at ₹199 a month.

Many Qwen models are open weights. The biggest, Qwen3.8-Max (August), has 2.4 trillion parameters with 95 billion active and a one-million-token context. Its licence allows commercial use, with conditions only for very large products and model-as-a-service businesses. Smaller models such as Qwen3.8-27B and Qwen3.6-35B-A3B are under Apache 2.0, and Qwen3.8-Omni-Flash understands text, images, audio and video. Watch the licences on image models: Qwen-Image-2.1 is research-only. Alibaba has said Qwen 4.5 and Qwen 5 are expected to scale up to between 5 and 10 trillion parameters; it has not published a size for Qwen 4.

4. Agents everywhere

ChatGPT in Chrome (15 October). OpenAI's browser extension, available on all ChatGPT plans, can look across your open tabs and gather information. In Work mode it can also control your Chrome tabs: fill out forms, send messages and download files. Slides also arrived for Pro, Business and Enterprise users, building a deck from a chat that your team can edit live and export to PowerPoint, PDF or Google Slides.

Manus. Manus raised more than $500 million on 8 October, its first round since its deal with Meta was unwound. On 13 October it announced that its agent will power a new Agent Mode in OPPO's AI Mind Pilot, starting with the Find X10 series, and launched team plans from $20 per seat per month.

What Manus 2.0 can do

Manus 2.0 launched on 28 September, and Manus says it "is not a version update" but a new architecture. In plain terms, Manus is an AI agent that does the work rather than describing it: research, slide decks, websites and apps, spreadsheets, video edits and simple games, on its own cloud computer. Version 2.0 adds four things worth knowing:

Automations: a task can now start by itself when a new email arrives, an ad's performance changes, a calendar event comes up, or a Slack or Notion update lands, not just at a scheduled time.
Manus Studio: the desktop app works directly with your local files and folders.
Remote Control: give a voice command from your phone and watch Manus operate your computer, using only the files and apps you approve.
Video Editor and Game Dev (1 October): Manus makes a first cut of a short video you can then edit on a timeline, or builds a playable game from a template that you can publish as a link.

Manus also launched Cue, a separate app in invite-only early access: "In Cue, each agent has its own email, phone number, wallet, and computer, so it can send messages, pay within the budget you set, and see a task through on its own machine." The free Manus plan gives 300 credits a day and one task at a time; Pro starts at $20 a month. Manus's iOS app is listed in both the Sri Lankan and Indian App Stores, though Manus doesn't publish a country list.

Notion, Claude and Microsoft. Notion 3.8 (13 October) connects its agents to more than 150 tools through MCP. Anthropic made Claude's Docs, Slides and Design tools free on every plan on 8 October and added Dashboards and Motion for paid plans; see our guide to Claude Dashboards and Motion. Microsoft said Copilot on Copilot+ PCs will be able to work with local files and take actions on the PC "in the coming months".

5. Voice, video and creative tools

Fish Audio Drama 3 (13 October): a text-to-speech model you direct in plain language, with tags like [barely holding it together], in 83 languages. The weights are on Hugging Face under a community licence that is free for commercial use below $500,000 in revenue and bans cloning a real person's voice without consent.
Kling 4.0 (15 October): 30-second video generation from one prompt, with up to 15 reference assets. Kuaishou lists Hindi among its languages and Indian English among its accents.
Sarvam Bulbul V4 Flash (13 October): an Indian-language voice model for real-time apps, under 250 milliseconds to the first audio, at ₹20 per 10,000 characters (₹10 at launch, until 12 November). No Sinhala yet.
ChatGPT Voice (15 October) moved to GPT-Live-1.1 and added new voices, including a Hindi one. Grok Voice 3.0 (14 October) scored 86.9 on the Artificial Analysis Speech-to-Speech Index, which xAI cites as first place.
ElevenLabs Instant Soundtracks (15 October) scores an uploaded video with original music and timed sound effects. ElevenLabs also voiced the UK Parliament's first AI-voiced speech, by Baroness Campbell, in a clone of her own voice: "It's my voice, my style, my words."
Google SynthID Detector (7 October) is open to everyone in English. Upload an image, video or audio file to check whether it was made with AI from Google or its partners. It is not a general AI detector. YouTube Music also started labelling songs made primarily with AI (14 October).

6. When AI agents went off-script

This was the fortnight AI agents' mistakes stopped being hypothetical. On 6 October OpenAI's Jason Kwon told Australia's parliamentary inquiry: "our models accessed Australian government websites in ways they were not directed to", and apologised. On 9 October Anthropic published its own report on Claude acting beyond its instructions on real websites; the New York Times reported that its agents filled out 20 visa applications on a US State Department site, none of which were processed.

Governments responded quickly. The White House told AI companies, in a statement given to Axios, that reporting such incidents "is not optional". India's CERT-In issued a critical advisory on 15 October after AI model providers disclosed that their models had "reached the live systems of Indian organizations". Singapore's IMDA released a starter kit for testing AI agents on 14 October.

If your team uses agents, the lesson is simple: give them the minimum access they need, keep a human approval step on anything that sends, submits, pays or deletes, and read their activity logs. Our free AI Policy builder is a good place to write those rules down.

7. Lawsuits and rules

On 14 October the attorneys general of Arizona, Arkansas, Idaho, Iowa, Kansas, Montana, Nebraska and West Virginia each sued OpenAI in their own state courts. The core claim is deceptive marketing: Iowa's attorney general says OpenAI has been "misleading the public into believing that ChatGPT is a safe tool that puts safety first". Harm to children is central to the suits. OpenAI has pointed to its age-specific protections. In Europe, the EU AI Office said it had sent requests for information to AI labs, which Euronews reported went to OpenAI and Anthropic.

8. Money and jobs

Twenty of the most valuable AI and semiconductor stocks lost more than $1 trillion in the week to 15 October, according to CNBC, with SK Hynix down more than 10% and Samsung down 8.73% in one session. The worries were the cost of the AI build-out and the chance of higher interest rates.

At the IMF and World Bank meetings in Bangkok on 12 October, IMF Managing Director Kristalina Georgieva called AI's effect on work "a slow-moving tsunami, but tsunami nonetheless", with about 40% of jobs worldwide set to be augmented, downgraded or eliminated and young graduates exposed first. Kroger announced plans to cut about 3,000 corporate jobs over three years. Money kept flowing into AI too: TypeSafe, the maker of the Jev model, raised about $870 million at a $7.5 billion valuation, Ramp raised $1.85 billion at $60 billion and Modal raised $750 million. Character.AI filed for Chapter 11 on 13 October.

9. Closer to home

Sri Lanka. On 12 October President Anura Kumara Dissanayake told the Ceylon Chamber's economic summit: "We expect to establish an AI center in Wijerama in collaboration with Singapore", with a Cabinet paper to follow. On 14 October Deputy Minister Eranga Weeraratne asked China's ambassador for support with a National Sovereign AI Computing Centre, and a draft memorandum on digital cooperation is being prepared.

Grok Bot Meetup Colombo. On Saturday 17 October the Grok Bot Meetup Colombo, hosted by Rasal Jayasinghe and Team TechTalk360 at the DHPL Auditorium, sold out both of its free sessions, with 467 people signed up on Luma. It was part of a community series running in more than 250 cities. We covered Grok Bot itself in September's AI agents roundup.

India. C-DAC opened an empanelment for up to 10,000 AI compute units for the IndiaAI programme on 7 October, and IT Minister Ashwini Vaishnaw promised a consultation paper on AI regulation "in a month's time" (it has not been published yet). Apple's HomeView smart display, with its new Siri AI, ships in India and seven other countries from 13 November.

Bangladesh and Pakistan. Bangladesh says it will launch sovereign AI infrastructure for Bangla language models within 12 months. Pakistan's law minister said AI-specific crime laws are being considered.

What this means for you

Try the free upgrades. GPT-6 and Claude's Docs, Slides and Design tools are now on free plans. Pick one real task this week and run it through both.
Don't pick a model from a headline. The leaderboards disagree, and the gaps are small. Test on your own work.
Put guardrails on agents before you need them. Map the workflow first with our free Workflow Mapper, then decide which steps an agent may do alone.
Check before you share. Deepfakes are getting easier to make. Use SynthID for Google-made media and our Scam Shield for suspicious messages.
Know where you stand. Take the free AI Readiness Score to see which of these changes you can use today.

Frequently asked questions

What were the biggest AI stories of early October 2026?

GPT-6 rolling out to every ChatGPT plan (7–8 October), Google's Gemini 4 Argon going global (13 October), Claude Haiku 5.5 (7 October), Meta's Muse Wonder and open-weight Muse Spark Open (12 and 15 October), AI agents acting on real government websites without being told to, eight US states suing OpenAI (14 October) and a sell-off that cost 20 of the biggest AI and chip stocks more than $1 trillion in a week.

Which AI model is the best right now?

It depends on the leaderboard. On Arena's public vote (15 October) Gemini 4 Argon leads, marked preliminary, just ahead of Muse Wonder and Claude Opus 5.5. On the Artificial Analysis Intelligence Index, Claude Opus 5.5 leads with 58. On the Vals Index, Gemini 4 Argon leads with 68.90%.

What AI news happened in Sri Lanka in October 2026?

On 12 October the President said Sri Lanka expects to set up an AI centre in Wijerama with Singapore's support, with a Cabinet paper to follow. On 14 October the Deputy Minister asked China's ambassador for help with a National Sovereign AI Computing Centre. On Saturday 17 October the Grok Bot Meetup Colombo sold out, with 467 people signed up on Luma.

Is it safe to let AI agents act on websites for me?

Be careful. In the first half of October OpenAI apologised to Australia's parliament after its models accessed government websites they were not directed to, Anthropic published a report on Claude acting beyond its instructions, and India's CERT-In warned that AI agents had reached the live systems of Indian organisations. Keep a human approval step on anything that sends, submits, pays or deletes.

What can Manus 2.0 do?

Manus 2.0 (28 September 2026) is an AI agent that carries out work such as research, slide decks, websites, spreadsheets, video edits and simple games on its own cloud computer. It can start tasks automatically when an email, calendar event or Slack or Notion update arrives, work with your local files through its desktop app, and operate your computer by voice through Remote Control. The free plan includes 300 credits a day.

What can Qwen do?

Alibaba's free Qwen app can search the web, write Deep Research reports, build web pages, and generate and edit images and videos. Many Qwen models are open weights, including Qwen3.8-Max with 2.4 trillion parameters. Qwen 4 is still in training as of 19 October 2026.

Sources

Leaderboards

New models

Open models

Agents

Voice, video and creative

AI agents off-script

Lawsuits and rules

Money and jobs

Sri Lanka, India and the region

Not sure which of these changes matter for your work? Find out where you stand in five minutes.

Take the free AI Readiness Score →